How to Choose AI Tools Without Buying Six of Them

The AI tool market is noisy, expensive, and full of products that are a form and a prompt wrapped around a model you can already access.

Some are genuinely valuable. Telling the difference takes about ten minutes per tool.

Try it in a general assistant first

The single most useful habit, and it saves most of the money.

Before buying a tool that does X, spend twenty minutes trying X in a general assistant you already pay for. Brief it properly — context, task, constraints, format, example — and see how close it gets.

Frequently it gets close enough, and the specialist tool was selling convenience you didn't need.

When it doesn't get close, you've learned something specific about what the tool would have to add, which makes evaluating it much faster.

What actually justifies a specialist tool

Not the model. Almost everyone uses the same handful of underlying models.

The four things that genuinely add value:

Integration. It's connected to your data and your other systems. A tool that reads your CRM and writes back to it is doing something a chat window can't.

Workflow. It handles a whole multi-step process rather than one step, including the boring parts around the edges.

Domain data. It's working from something you don't have — a specialist corpus, current market data, a body of regulations.

Volume and interface. It processes a hundred items with a review queue rather than one at a time in a text box.

If a tool doesn't do at least one of those, you're paying a margin on something you can do directly.

Check what you already pay for

Most major business software has added AI features. Before buying anything new:

• Your accounting software — categorisation, anomaly detection
• Your CRM — drafting, summarising, next-step suggestions
• Your email and documents — drafting, summarising
• Your support tool — reply drafting, tagging, sentiment
• Your project tool — summarising, updates

This is the same discipline as choosing any automation tool: work down from what you already own before adding a dependency. A large share of AI tool spending replicates a feature already included in something the business pays for.

The questions before subscribing

Six, and they take ten minutes:

What happens to what I put in? Is it used for training, who can access it, where is it stored. This varies by plan within the same product — check the tier you'd actually buy.

How do I get my data out? If your workflow ends up depending on this, export format is the difference between switching and being stuck.

What does it cost at three times current usage? Per-seat and per-use pricing scale very differently. Growth shouldn't be punished by your tool choice.

What happens when the output is wrong? Is there a review step, an audit trail, a way to see what it did and why.

Who maintains this if the person who set it up leaves?

Is this vendor likely to exist in two years? This market consolidates and shuts down quickly. Prefer tools where being wrong about longevity is cheap.

Trial it properly

A fortnight, on real work, with one honest number: time saved after verification.

Not time saved. Time saved minus the time spent checking, correcting and reformatting the output. That number is frequently a third of what the marketing claims, and it's still often worth it — but you should decide with the real figure.

Also track:

• How often the output needed material correction
• Whether anyone other than you would use it — a tool only you understand is a dependency, not a system
• What broke when it hit something unusual

Then decide. Most tools that survive a fortnight of real use are worth keeping; most that don't were bought on enthusiasm.

Audit quarterly

AI subscriptions accumulate faster than most, because they're individually cheap and easy to justify.

Once a quarter, list every tool with what it costs, what it does, and who used it in the last month. Two questions:

• Is anything doing a job another tool now includes? Features get absorbed into platforms constantly in this market.
• Has anyone opened this in three months?

Cancel accordingly. This is exactly the automation debt problem in a different costume — unused subscriptions and half-configured tools that nobody owns.

What not to optimise for

The newest model. Model quality changes constantly and the differences at the top matter far less for ordinary business tasks than the briefing does. Choosing a tool because of which model it uses this month is optimising the wrong variable.

Feature lists. Most go unused. One workflow that works beats forty features.

Demos. Demos are built on ideal inputs. Trial on your actual messy real work, which is where tools fail.

The mistakes

1. Buying before trying the task in a general assistant. The cheapest test available.
2. Not checking what your existing software already includes.
3. Choosing on which model it uses. Optimising the wrong variable.
4. Trialling on demo data. Real work is where tools break.
5. Counting time saved without verification time. Overstates by roughly a third.
6. Never auditing. Six subscriptions, two in use.

What to do next

List every AI tool you currently pay for, with the monthly cost and who last used it. Then for the most expensive one, spend twenty minutes trying its core task in a general assistant with a proper brief.

If the general assistant gets close, you've found a cancellation. If it doesn't, you now know exactly what you're paying for — which is a better position than assuming either way.